Energy-Ordered Prequential Tracking of the Muon Puzzle

Every cosmic-ray observatory interprets its data through simulations of how particles collide in the atmosphere, and those simulations are known to be wrong. They predict too few muons, and how wrong they are seems to change with energy. We propose to treat the competing simulation packages as "experts" whose relative merit and required correction are allowed to drift as one moves up in energy. We then fit that drifting correction with online learning algorithms that carry dynamic-regret guarantees, meaning their predictive loss is provably close to that of the best drifting sequence of corrections chosen in hindsight. The scientific output is a correction profile and a single number, the total amount of energy-dependent change the data actually require. Because the guarantees hold without choosing a smoothing level in advance, the method also says how much that answer depends on the analyst's choices.

The stakes are concrete. Cosmic rays above $10^{17}$ eV hit the atmosphere at collision energies far above the Large Hadron Collider's roughly 13–14 TeV. Their showers are simulated with hadronic interaction models (event generators such as QGSJET-II-04, EPOS-LHC and SIBYLL 2.3d that predict which particles a nuclear collision produces), and these models must extrapolate into untested territory. Telescope Array finds a ground muon signal at 2000 m from the shower core that is $1.67 \pm 0.10 \pm 0.36$ times the QGSJET-II-04 proton prediction, with the ratio exceeding 3 at 3000 m (Study of muons… Telescope Array). Auger inclined showers exceed predictions by 23% to over 100% (The Pierre Auger Observatory I). At PeV energies, LHAASO's muon attenuation length (449 to 811 g/cm² from 0.3 to 30 PeV) sits close to EPOS-LHC but well short of the other two models (LHAASO muon attenuation). Composition results inherit the problem: the Telescope Array surface-detector composition fit gives $\langle \ln A\rangle = 2.0 \pm 0.1(\text{stat}) \pm 0.44(\text{syst})$, and the systematic part is dominated by the choice of hadronic model (TA SD mass composition).

Core Claim

The open question is not whether models are off. It is how their error evolves across energy, and in particular how many distinct changes in hadronic behaviour the data actually support. Current practice answers this either with one rescaling per energy bin or with an assumed functional form. Both choices bake in the very complexity that should be measured.

Our claim is that a prequential, energy-ordered online learner answers this question in a way existing practice does not. "Prequential" means scoring each event by how well it was predicted from events already seen. "Dynamic regret" means comparing the learner to the best sequence of corrections that is allowed to change from event to event, rather than to a single fixed correction. Such a learner provides two things. The first is a correction profile whose excess predictive loss is bounded in terms of the unknown total variation of the true correction. The second is a cumulative predictive score that penalises unnecessary drift automatically, with no cross-validation over a smoothing parameter. Cross-validation is fragile at the highest energies, where events are scarce.

The analogy to control theory is exact rather than decorative. Nonasymptotic regret analysis of adaptive LQ control with model misspecification shows that a learner restricted to a fixed, misspecified basis of models adapts quickly but then accumulates regret linearly, roughly $\delta T$ for misspecification $\delta$. A static choice of hadronic generator is that fixed basis. The remedy we test is to let the comparator move, not to discard the models.

Method

Events are sorted by reconstructed energy $E$, which becomes the "time" axis $t = 1,\dots,T$. Each event $t$ has observables $y_t$: the muon-dominated surface signal at a fixed distance and, for hybrid events, the depth of shower maximum $X_{\max}$ measured by fluorescence telescopes.

We build a finite dictionary of cells $k = (m, A, r)$. Here $m$ is the generator, $A$ is the primary mass group (p, He, N, Fe), and $r$ is a muon rescaling factor on a grid from 0.8 to 3.5. Each cell supplies a simulated likelihood $p_k(y \mid E)$ obtained from CORSIKA showers passed through the detector simulation.

The learner predicts each new event with a mixture $\sum_k w_{t,k}\, p_k(y_t \mid E_t)$ and pays the log-loss

$$\ell_t(w) = -\log \sum_k w_k\, p_k(y_t \mid E_t).$$

This loss is exp-concave in the weights $w$, which roughly means its curvature can be exploited for faster learning. Optimal Dynamic Regret in LQR Control shows that exploiting exactly this kind of curvature is what yields the optimal $\tilde O(T^{1/3}\mathcal{TV}^{2/3})$ dynamic-regret rate and $O(\log|I|)$ regret on every sub-interval. We reuse that curvature-preserving approach, through strongly adaptive Online Newton Step and fixed-share mixtures, on the simplex of cells, where projection is easy. Baby and Wang need known dynamics to obtain this structure. Our problem has it natively, because the "dynamics" are just the energy ordering.

The guarantee we target bounds cumulative log-loss relative to any comparator sequence $w^\star_{1:T}$ by a term growing with its path length $\sum_t \|w^\star_{t+1} - w^\star_t\|_1$. That path length has a physical reading: it is the total change in composition and in hadronic correction across energy. A minimum-description-length dual follows directly. The drop in prequential code length when switching is allowed, compared with a static mixture, measures how much drift the data demand.

Composition and hadronic physics are degenerate: both can shift the muon signal. Cells therefore share the same $A$ across generators. Hybrid $X_{\max}$, which depends far less on the muon problem, anchors the composition weights, while the surface muon signal constrains $r$. This mirrors the Auger golden-hybrid logic of anchoring surface data to near-calorimetric fluorescence measurements (Auger I).

Research Program

The work has three parts.

The first is the dictionary. We generate CORSIKA libraries for the three generators and four mass groups across $10^{18}$–$10^{20}$ eV for a Telescope-Array-like scintillator array, and across 0.3–30 PeV for a LHAASO-like muon detector. Rescaling is implemented as event-level reweighting of the muon component, so the $r$ grid costs no extra simulation.

The second is theory specialised to this setting. We prove dynamic-regret bounds for the log-loss mixture in which the comparator is constrained to cells sharing a common $A$-marginal. We also bound how the misspecification floor of the misspecified-LQ analysis transfers: a static comparator should pay linear excess loss whenever the true $r(E)$ varies.

The third is application, first to mock data with injected truth, then to real events where event-level access is available. Otherwise we use mock datasets tuned to the published summaries above.

The energy ordering is a modelling choice: events of similar energy are exchangeable. We therefore report results for ascending order, descending order, and random within-bin permutations. A conclusion that flips with the order is not reported as physics.

Baselines

The naive reference is the best single static cell set: one generator, one rescaling, and composition fractions fit once on all data. This is the "do nothing" analysis. Under drifting truth it should show the linear misspecification floor.

The standard practice baseline fits rescaling and composition independently in each energy bin, as in the muon-excess analyses of Telescope Array. Its reported anchor is a muon ratio of $1.67 \pm 0.10 \pm 0.36$ for QGSJET-II-04 at 2000 m. We compare against it on the same events and with the same simulation library.

The composition-side reference is the 14-observable boosted-decision-tree analysis of TA SD composition. Its hadronic systematic of 0.44 on $\langle\ln A\rangle$ is the number to reduce. The test is the spread of inferred $\langle \ln A\rangle$ across generators, once each is allowed its own data-determined drifting correction.

The strongest competitor is offline, not online. It is a trend-filtering or hierarchical-Bayes fit of $r(E)$ and composition fractions with a total-variation or Gaussian-process smoothness prior, with the hyperparameter set by cross-validation or marginal likelihood. No reported result was found on Lacuna for this baseline in the cosmic-ray setting, so we will establish it ourselves. It gets the identical dictionary, likelihoods, data and an equal compute budget for hyperparameter search.

Two ablations isolate the ingredients. The first removes switching, giving a static online mixture that achieves only static regret. The second removes the $X_{\max}$ anchor, which tests whether degeneracy breaking comes from the shared composition structure.

Beating the strongest competitor means one of three outcomes on injected-truth mocks: lower integrated error on $r(E)$, more accurate recovery of the number and location of injected breaks, or a lower generator-to-generator spread in $\langle\ln A\rangle$. Each must hold at matched event counts and especially in the sparse top energy decade.

Why not just…?

Why not just re-tune the generators on new collider data? Post-LHC tunes already exist, and the excess persists against them: 1.67 relative to QGSJET-II-04 at Telescope Array, and 23% to over 100% at Auger. Colliders cannot reach the relevant collision energies. Re-tuning is complementary, and our correction profile tells re-tuners where in energy the extrapolation must change.

Why not just fit each energy bin separately? This is current practice. It is unbiased but noisy where events are rare, and it hard-codes bin width as the implicit complexity. In mocks we will show whether it over-reports structure at the highest energies relative to the prequential learner. If it does not, the online machinery adds little for this observable, and we will say so.

Why not just use offline trend filtering or a hierarchical Bayesian model? This is the real threat. On a fixed archived dataset, a well-regularised offline fit may estimate $r(E)$ as accurately as our learner. Our claimed advantage is narrower. The prequential score selects the amount of drift without a tuned hyperparameter, and the guarantee holds for every sub-interval simultaneously, which matters when the top decade has too few events for stable cross-validation. The mock experiments with injected step and smooth drifts at realistic event counts settle this directly. If offline trend filtering with cross-validation matches us on accuracy and break recovery, we retain only the drift-budget statistic and the streaming use case.

Why not just learn shower physics model-free from data? The highest-energy events are far too rare. Both the gray-box analysis of He and Muehlebach and the misspecified-LQ result show that model priors give a large early advantage. Our dictionary keeps that advantage and relaxes only the fixed-comparator assumption.

Proposed Experiments

In the decisive mock test, we inject an energy-dependent muon rescaling that rises from 1.2 to 1.7 across $10^{18.5}$–$10^{19.2}$ eV, matching the published Telescope Array scale, on top of a composition that shifts from light to intermediate. We then compare the methods on recovery of $r(E)$, of composition, and of the injected path length.

A second test injects a single abrupt break against a smooth alternative and asks which method's selection criterion recovers the truth at 2,000, 5,000 and 18,000 events. The last count matches the TA composition sample.

A third test uses LHAASO-like PeV mocks in which EPOS-LHC is nearly correct. There, the learner should converge to near-zero drift and concentrate weight on EPOS-LHC. This is a falsifiable check that it does not invent structure.

On real or published-summary data, we report the required drift budget separately for PeV and EeV regimes. We then ask whether one generator with a modest, smoothly growing correction can account for both the LHAASO attenuation lengths and the Telescope Array muon excess, or whether the data demand a sharper change between them. That is the quantitative form of the tension the field currently describes only qualitatively.